search

huggingface / distil-whisper

Distilled variant of Whisper for speech recognition. 6x faster, 50% smaller, within 1% word error rate.
MIT License
3.33k stars 238 forks source link

Training README triggers BuildConfig ValueError #86

Closed guynich closed 4 months ago

guynich commented 4 months ago

I'm following the training README - it is very well documented thanks!

I'm seeing a ValueError with the Stage 3 training script. I documented all my steps in this README and show the error here https://github.com/guynich/distil-whisper/blob/main/training/scripts/README.md#training-error-on-ubuntu-with-a10-gpu. The paths to my student model and pseudo-labelled folder both look correct,

The error is ValueError: BuilderConfig 'hi' not found. Available: ['default']. My run_pseudo_labelling script uses the training README dataset_config_name "hi" here.

Any advice for a mitigation please? Thanks.

guynich commented 4 months ago

This error does not duplicate on another workstation with a different GPU. Which suggests some dependency issues. I have updated the above README with more information.

guynich commented 4 months ago

The error is ValueError: BuilderConfig 'hi' not found. Available: ['default'].

Changing the run distillation script train_dataset_config_name="default+default" and eval_dataset_config_name="default" mitigates the error however having run the pseudo-labelling with config names of "hi+hi" and "hi" then running training with different config names may causes other problems and seems "hacky" ?

guynich commented 4 months ago

I resolved this error by changes to the environment documented here. Closing.